Deep subspace anomaly detection for radar target detection in heterogeneous clutter environments

Radar target detection in heterogeneous clutter environments represents a critical challenge in signal processing. Conventional detection methods, which rely on the assumption of local stationarity and fixed statistical models, suffer from severe performance degradation in such complex scenarios. To address model mismatch issues caused by spatial heterogeneity in the statistical properties of noise, we have proposed a framework based on deep subspace detection. This method abandons the traditional approach of globally modeling clutter uniformly. Instead, it constructs a parallel array of deep subspace autoencoders to adaptively decompose high-dimensional radar data into multiple low-dimensional subspaces in an unsupervised manner. This enables the precise characterization of diverse local statistical patterns within heterogeneous clutter. Furthermore, we designed an anomaly scoring mechanism that integrates multi-scale evidence. By jointly analyzing subspace reconstruction error, subspace consistency, local neighborhood isolation, and temporal dynamic characteristics, we constructed a robust and highly discriminative anomaly indicator. Finally, we employed an adaptive threshold strategy based on data distribution to achieve a constant false alarm rate detection. Experiments conducted on complex simulation scenarios featuring multiple statistical distribution types of clutter, discrete interference, and fluctuating targets demonstrated that the proposed method significantly outperforms advanced data-driven approaches. This study presents an effective data-driven solution for robust radar target detection in complex electromagnetic environments that does not rely on precise prior knowledge.